V11 specialist LoRA adapter
This is the deployed LegalChatModel adapter for
mlx-community/Qwen2.5-7B-Instruct-Uncensored-4bit.
- Fine-tuning type: LoRA, rank 8, 16 layers
- Context used for training: 4,096 tokens
- Selected validation: V11 outperformed the preceding V10 adapter on the
untouched test set (loss 2.681 vs 2.698; perplexity 14.601 vs 14.844)
- Privacy audit: the model-visible corrective datasets contain no local paths,
candidate identifiers, private filenames, chat transcripts, or lower-mark
submitted prose
Only the final adapters.safetensors and its configuration are published.
Intermediate checkpoints and the rejected V12 experiment are intentionally
excluded.
An adapter is not a legal authority or an accuracy guarantee. Runtime official
source checking, authority repair, privacy filtering, and release supervision
remain necessary.
